Exploiting Sentence Similarities for Better Alignments

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2193–2203, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Exploiting Sentence Similarities for Better Alignments Tao Li and Vivek Srikumar University of Utah {tli,svivek}@cs.utah.edu Abstract We study the problem of jointly aligning sen- tence constituents and predicting their sim- ilarities. While extensive sentence similar- ity data exists, manually generating reference alignments and labeling the similarities of the aligned chunks is comparatively onerous. This prompts the natural question of whether we can exploit easy-to-create sentence level data to train better aligners. In this paper, we present a model that learns to jointly align constituents of two sentences and also predict their similarities. By taking advantage of both sentence and constituent level data, we show that our model achieves state-of-the-art per- formance at predicting alignments and con- stituent similarities. 1 Introduction The problem of discovering semantic relationships between two sentences has given birth to several NLP tasks over the years. Textual entailment (Da- gan et al., 2013, inter alia) asks about the truth of a hypothesis sentence given another sentence (or more generally a paragraph). Paraphrase identifi- cation (Dolan et al., 2004, inter alia) asks whether two sentences have the same meaning. Foregoing the binary entailment and paraphrase decisions, the semantic textual similarity (STS) task (Agirre et al., 2012) asks for a numeric measure of semantic equiv- alence between two sentences. All three tasks have attracted much interest in the form of shared tasks. While various approaches have been proposed to predict these sentence relationships, a commonly employed strategy (Das and Smith, 2009; Chang et Gunmen abduct 7 foreign workers ∅ Seven foreign workers kidnapped EQUI EQUI Figure 1: Example constituent alignment. The solid lines represent aligned constituents (here, both la- beled equivalent). The chunk Gunmen is unaligned. al., 2010a) is to postulate an alignment between con- stituents of the sentences and use this alignment to make the final prediction (a binary decision or a nu- meric similarity score). The implicit assumption in such approaches is that better constituent alignments can lead to better identification of semantic relation- ships between sentences. Constituent alignments serve two purposes. First, they act as an intermediate representation for pre- dicting the final output. Second, the alignments help interpret (and debug) decisions made by the over- all system. For example, the alignment between the sentences in Figure 1 can not only be useful to deter- mine the equivalence of the two sentences, but also help reason about the predictions. The importance of this intermediate representa- tion led to the creation of the interpretable seman- tic textual similarity task (Agirre et al., 2015a) that focuses on predicting chunk-level alignments and similarities. However, while extensive resources ex- ist for sentence-level relationships, human annotated chunk-aligned data is comparatively smaller. In this paper, we address the following question: can we use sentence-level resources to better pre- 2193 dict constituent alignments and similarities? To an- swer this question, we focus on the semantic tex- tual similarity (STS) …

Content truncated to 500 words